Machine learning integrating CT radiomics and clinical variables for preoperative prediction of renal uric acid stones: a model development and validation study.
Authors
Affiliations (5)
Affiliations (5)
- Department of Urology, The Fourth Affiliated Hospital of Soochow University, Suzhou, China.
- Department of Urology, The Third People's Hospital of Hefei, Hefei, China.
- Department of Neurosurgery, The First Affiliated Hospital of Soochow University, Suzhou, China.
- Department of Urology, The First Affiliated Hospital of Soochow University, Suzhou, China.
- The Second Clinical College of Anhui Medical University, Hefei, China.
Abstract
Uric acid stones may be amenable to pharmacological dissolution, making accurate preoperative identification important for targeted management in selected patients. Existing methods, including Hounsfield unit (HU) thresholds on non-contrast computed tomography (CT) and urine pH criteria, may be insufficient when used alone, while dual-energy CT is limited by cost and availability. This study aimed to develop and internally validate machine learning models integrating CT radiomics with clinical parameters for non-invasive preoperative prediction of uric acid stone composition. We conducted a retrospective prediction model development and validation study of 371 patients who underwent percutaneous nephrolithotomy (PCNL) for renal stones at our institution between May 2019 and July 2024. Eligible patients had preoperative non-contrast CT images, PCNL stone removal, and postoperative infrared spectroscopy-based composition analysis. Patients with incomplete clinical information, prior ipsilateral lithotripsy, or use of uric acid-lowering or lipid-lowering medications were excluded. Patients were divided into training (n=296) and test (n=75) cohorts at an 8:2 ratio. Clinical predictors were identified using univariate and multivariate logistic regression. Radiomics features were extracted from preoperative non-contrast CT images using PyRadiomics and selected by variance analysis and least absolute shrinkage and selection operator (LASSO) regression. Four machine learning classifiers, CatBoost, XGBoost, decision tree (DT), and random forest (RF), were trained using radiomics features alone and in combination with clinical variables. Model performance was evaluated using five-fold cross-validation and metrics including area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. Of 371 patients, 60 (16.2%) had uric acid stones and 311 (83.8%) had non-uric acid stones. Multivariate logistic regression identified age [odds ratio (OR) =1.063, 95% confidence interval (CI): 1.027-1.099, P<0.001], urine pH (OR =0.106, 95% CI: 0.047-0.235, P<0.001), serum uric acid (OR =1.005, 95% CI: 1.002-1.009, P=0.003), and high-density lipoprotein (HDL) cholesterol (OR =0.041, 95% CI: 0.008-0.212, P<0.001) as independent predictors of uric acid stone formation. Ten radiomics features were selected from 1,316 extracted features. Radiomics-only models achieved AUC values of 0.980-0.983 in the test set. Integrated clinical-radiomics models showed high predictive performance, with the CatBoost model achieving the highest test-set AUC of 0.991 (accuracy =0.933, sensitivity =0.917, specificity =0.937). Machine learning models combining CT radiomics features with clinical variables demonstrated promising internal validation performance for the non-invasive preoperative prediction of renal uric acid stones. Age, urine pH, serum uric acid, and HDL cholesterol were identified as independent predictors. The CatBoost integrated model showed the highest discrimination; however, external prospective multicenter validation is required before clinical application.